NERDA

Not only is NERDA a mesmerizing muppet-like character. NERDA is also a python package, that offers a slick easy-to-use interface for fine-tuning pretrained transformers for Named Entity Recognition (=NER) tasks.

You can also utilize NERDA to access a selection of precooked NERDA models, that you can use right off the shelf for NER tasks.

NERDA is built on huggingface transformers and the popular pytorch framework.

Installation guide

NERDA can be installed from PyPI with

pip install NERDA

If you want the development version then install directly from GitHub.

Named-Entity Recogntion tasks

Named-entity recognition (NER) (also known as (named) entity identification, entity chunking, and entity extraction) is a subtask of information extraction that seeks to locate and classify named entities mentioned in unstructured text into pre-defined categories such as person names, organizations, locations, medical codes, time expressions, quantities, monetary values, percentages, etc.¹

Example Task:

Task

Identify person names and organizations in text:

Jim bought 300 shares of Acme Corp.

Solution

Read more about NER on Wikipedia.

Train Your Own NERDA Model

Say, we want to fine-tune a pretrained Multilingual BERT transformer for NER in English.

Load package.

from NERDA.models import NERDA

Instantiate a NERDA model (with default settings) for the CoNLL-2003 English NER data set.

from NERDA.datasets import get_conll_data
model = NERDA(dataset_training = get_conll_data('train'),
              dataset_validation = get_conll_data('valid'),
              transformer = 'bert-base-multilingual-uncased')

By default the network architecture is analogous to that of the models in Hvingelby et al. 2020.

The model can then be trained/fine-tuned by invoking the train method, e.g.

model.train()

Note: this will take some time depending on the dimensions of your machine (if you want to skip training, you can go ahead and use one of the models, that we have already precooked for you in stead).

After the model has been trained, the model can be used for predicting named entities in new texts.

# text to identify named entities in.
text = 'Old MacDonald had a farm'
model.predict_text(text)
([['Old', 'MacDonald', 'had', 'a', 'farm']], [['B-PER', 'I-PER', 'O', 'O', 'O']])

This means, that the model identified 'Old MacDonald' as a PERson.

Please note, that the NERDA model configuration above was instantiated with all default settings. You can however customize your NERDA model in a lot of ways:

• Use your own data set (finetune a transformer for any given language)
• Choose whatever transformer you like
• Set all of the hyperparameters for the model
• You can even apply your own Network Architecture

Read more about advanced usage of NERDA in the detailed documentation.

Use a Precooked NERDA model

We have precooked a number of NERDA models for Danish and English, that you can download and use right off the shelf.

Here is an example.

Instantiate a multilingual BERT model, that has been finetuned for NER in Danish, DA_BERT_ML.

from NERDA.precooked import DA_BERT_ML()
model = DA_BERT_ML()

Down(load) network from web:

model.download_network()
model.load_network()

You can now predict named entities in new (Danish) texts

# (Danish) text to identify named entities in:
# 'Jens Hansen har en bondegård' = 'Old MacDonald had a farm'
text = 'Jens Hansen har en bondegård'
model.predict_text(text)
([['Jens', 'Hansen', 'har', 'en', 'bondegård']], [['B-PER', 'I-PER', 'O', 'O', 'O']])

List of Precooked Models

The table below shows the precooked NERDA models publicly available for download.

F1-score is the micro-averaged F1-score across entity tags and is evaluated on the respective test sets (that have not been used for training nor validation of the models).

Note, that we have not spent a lot of time on actually fine-tuning the models, so there could be room for improvement. If you are able to improve the models, we will be happy to hear from you and include your NERDA model.

Model Performance

The table below summarizes the performance (F1-scores) of the precooked NERDA models.

'NERDA'?

'NERDA' originally stands for 'Named Entity Recognition for DAnish'. However, this is somewhat misleading, since the functionality is no longer limited to Danish. On the contrary it generalizes to all other languages, i.e. NERDA supports fine-tuning of transformers for NER tasks for any arbitrary language.

Background

NERDA is developed as a part of Ekstra Bladet’s activities on Platform Intelligence in News (PIN). PIN is an industrial research project that is carried out in collaboration between the Technical University of Denmark, University of Copenhagen and Copenhagen Business School with funding from Innovation Fund Denmark. The project runs from 2020-2023 and develops recommender systems and natural language processing systems geared for news publishing, some of which are open sourced like NERDA.

Shout-outs

• Thanks to Alexandra Institute for with the danlp package to have encouraged us to develop this package.
• Thanks to Malte Højmark-Bertelsen and Kasper Junge for giving feedback on NERDA.

Read more

The detailed documentation for NERDA including code references and extended workflow examples can be accessed here.

Contact

We hope, that you will find NERDA useful.

Please direct any questions and feedbacks to us!

If you want to contribute (which we encourage you to), open a PR.

If you encounter a bug or want to suggest an enhancement, please open an issue.